Neural Network Classification Results

Information and Communication Technology Seminar, Vol. 1 No. 1, August 2005 ISSN 1858-1633 2005 ICTS 32 Figure 1. Results of Land Cover Classification in Subset of the Study Area Using Mahalanobis distance algorithm, the fuzzy c-means was more aggressive to classify single tree felling STF class resulting higher classification accuracy of single tree felling class compared to the Euclidian distance algorithm. In general, classification accuracy of fuzzy c- means using Euclidian distance is slightly higher than the accuracy of Mahalanobis distance.

4.2. Neural Network Classification Results

In the following analysis, Landsat ETM data was used as an input for neural network classification and single hidden layer architecture was applied. Kanellopoulos [25] in his study has found that the use of a single hidden layer was sufficient for most classification problems, however, once the number of inputs gets near 20, additional flexibility was required as provided by a two hidden layer network. Total system Root Mean Squared RMS error of 0.0001 was determined as a convergence point. Training was stopped when convergence was reached, or the network reached an asymptote point when training accuracy started decreasing. By default, neural network application used the equal number of hidden nodes as the number of input variable. Skidmore, et al. [28] found that the use of minimum number of hidden nodes in the neural network significantly reduced the average training accuracy, resulting in a lower accuracy of the classification result. His study found that mean training accuracy increased as more hidden nodes were added. Another study mentioned that it was sometimes useful to make the number of hidden nodes roughly equal to two or three times the total number of input classes [25]. This study used two variations of hidden nodes number, which are equal and three times of the total input number used in the neural network, while holding other parameters constant. Analysis on the classification results found that the use of more hidden nodes number in the neural network made the network architecture more complex, causing more complicated computation for training the network, which in turn needed more iterations to reach global minima. As a comparison, neural network with 7 hidden nodes reached convergence point after 5,000 iterations, whereas the use of 21 hidden nodes in the network resulted in longer training of 7,500 iterations in order to generate a similar training accuracy. Neural network was trained using back- propagation learning algorithm with learning rate and Application of Soft Classification Techniques for Forest Cover Mapping – Arief Wijaya ISSN 1858-1633 2005 ICTS 33 momentum value of 0.2 and 0.4, respectively. Learning rate reflects on the training speed, while momentum describes the sensitivity of the network to error surface. This study tried to use some variations on these parameters, and found that higher learning rate value should be balanced with the higher value of momentum, otherwise training stage became unstable and was trapped into local minima condition. According to the accuracy assessment on classification results, the best performance of neural network was achieved with 21 hidden nodes when the network was trained for 7,500 iterations.

4.3. Comparison of Classification Results